Is a protein's trapping written in its starting structure?

Single protein molecules often fail to explore their full set of shapes on biologically relevant timescales. In smFRET experiments (single-molecule fluorescence measurements that track the distance between two dyes on one molecule), some copies of the phosphatase SHP2 stay stuck in one conformation for 100 seconds while others visit two or three. Glass physics has a sharp tool that biophysics has never applied to this problem: the isoconfigurational ensemble. The idea is to restart many simulations from the identical atomic structure with different random initial velocities. Averaging over those restarts separates what the structure itself determines from what thermal noise determines. In supercooled liquids, this tool showed that a particle's future mobility is partly encoded in a snapshot that looks featureless. That is the problem of "structural-dynamic decoupling": the liquid's structure barely changes while its dynamics slow by many orders of magnitude.

We propose to apply this test to proteins. The question is whether molecule-to-molecule differences in long-time behavior are encoded in the starting structure at a level finer than which coarse conformational basin the molecule starts in. We also ask whether that encoding changes with the molecule's age since preparation. If the answer is yes, it yields a measurable "structural fraction" of non-ergodicity, a predictor for which molecules will be trapped, and a new way to argue about whether protein aging is real or a sampling artifact.

Core Claim

The spread of time-averaged observables across molecules can be split into two parts. One part comes from the starting configuration; the other comes from the thermal history that follows. For the time-averaged value $\bar A_{ij}(T)$ of an observable over window $T$, where the trajectory starts from configuration $i$ with velocity seed $j$, define the structural fraction

$$ \phi(T) = \frac{\mathrm{Var}_i\big(\mathbb{E}_j[\bar A_{ij}(T)]\big)}{\mathrm{Var}_{ij}\big(\bar A_{ij}(T)\big)} . $$

The numerator is the variance, across starting configurations, of the average over velocity seeds. The denominator is the total variance over all trajectories. For an ergodic system, both variances shrink as $T$ grows and $\phi$ carries little information. For a system whose trapping is set by structure, $\phi$ stays high over the whole window.

The claim is threefold. First, globular proteins with branched, fractal-like landscapes will show large $\phi$ over windows where time averages have not converged. Second, a comparably sized ssDNA will not show this. Third, a meaningful part of $\phi$ survives after we condition on the coarse starting basin, and a local geometric predictor can recover that part.

Prior work motivates this claim without testing it. Non-ergodicity of a globular protein extending beyond its functional timescale ran 100 independent 100-ns simulations of SHP2. It found that time-averaged and ensemble-averaged mean-squared displacement disagree, with the time-averaged value aging as $t^{-0.2}$. It attributed this to a transition network with fractal dimension about 1.7, against about 1.0 for ssDNA. Its trajectories started from independent states, so it cannot say whether the divergence is structurally encoded. The dynamics of single protein molecules is non-equilibrium and self-similar over thirteen decades in time models the same phenomenon as a continuous-time random walk: a walker waits a random, heavy-tailed time in each trap before hopping. That paper finds relaxation time scaling as $t^{0.9}$. Its model treats trap depths as random draws with no spatial or structural content. Our test asks whether those "random" trap depths are in fact predictable from atomic geometry.

Research Program

The method comes from glass work. Structure–dynamics correlation and its link to fragility and dynamic heterogeneity used 100 isoconfigurational runs per configuration in Kob–Andersen mixtures to measure structure–mobility correlation. It also constrains our expectations. It found that the correlation depends on enthalpy dominance and nearness to mechanical instability, not on fragility (how sharply dynamics slow on cooling). So there is no reason to expect a large $\phi$ simply because a protein is "glassy." Enhancing the Prediction of Glass Dynamics by Incorporating the Direction of Deviation from Equilibrium Positions supplies the predictor design. Its model, EIGNN, is a graph neural network, and its key input is the direction from each particle to its vibrational cage center, not just the distance. Removing that direction sharply hurt performance.

For proteins, we define the cage center of each heavy atom (or residue centroid) as its mean position over a short restrained run of a few picoseconds. The network then receives vector displacements to these centers on a residue contact graph.

Work proceeds in three stages. First, we validate the pipeline on GlassBench-3D, a public benchmark of the Kob–Andersen binary Lennard-Jones liquid with mobility labels from repeated runs. We compute $\phi$ for particle displacement at $T=0.44$ and $T=0.64$. We expect to reproduce the known picture: strong structural encoding at low temperature, weaker at high temperature, where frequent hops break the cage-center reference.

Second, we build isoconfigurational ensembles for SHP2 and for the ssDNA control, reproducing the published simulation protocol. We draw 50 starting configurations, each restarted with 20 velocity seeds and run for 100 ns, for about 100 μs of simulation per system. Starting configurations are drawn at several waiting times $t_w$ along long reference trajectories, so that $\phi$ can be measured as a function of age. We then extend to PGK, a two-domain protein, and K-Ras, a single-domain protein, following the protocol of the thirteen-decades study. Where those Anton trajectories are unavailable, we rerun comparable systems. The ATLAS set of about 1,300 100-ns monomer simulations is used only to screen candidate proteins by how far their time-averaged and ensemble-averaged MSD disagree.

Third, we train predictors that map a single starting structure to its seed-averaged time-averaged observable. Two observables are used: the time-averaged MSD, and an inter-domain distance matching the smFRET coordinate. We then ask how much of $\phi$ each predictor recovers.

Varying the predictor's receptive field (the graph radius it may see) gives a secondary readout. That readout bears on the dispute in Local vs. Cooperative: Unraveling Glass Transition Mechanisms with SEER, which found that local barriers alone explain slowdown in soft spheres, with fewer particles participating on cooling. If trapping predictability in proteins saturates at a small radius, trapping is local, for example set by salt bridges. If predictability requires seeing across domain interfaces, trapping is cooperative.

Baselines

The naive reference predicts every configuration's seed-averaged outcome as the population mean. This reference recovers none of $\phi$, and every other method is scored by the fraction of between-configuration variance it explains on held-out starting structures.

The first real competitor, and the one we expect to be hardest to beat, is the starting-basin label. Each start is assigned to its conformational cluster or FRET-like state, using the clustering of the SHP2 non-ergodicity study, and the prediction is that cluster's mean outcome. This baseline matters because the SHP2 smFRET data already show molecules trapped in one of four states. It is quite plausible that "which state you start in" explains most of the spread. It uses the same held-out split and a cluster count tuned by cross-validation. Beating it means explaining a statistically significant further share of $\phi$ within basins, with confidence intervals estimated by bootstrapping over starting configurations.

The second baseline is a static-landscape kinetic model built at the finest resolution the data support. It is either a Markov state model or the PESLA approach from Predicting the Energy Landscape of Stochastic Dynamical System via Physics-informed Self-supervised Learning. PESLA groups states with a learned codebook and fits a graph-based drift-diffusion model of transitions between them. It reports an absolute Pearson correlation above 0.9 with ground-truth energy and an average 17.65% improvement in future-state prediction over NeuralMJP and SDE-Net on fast-folding proteins. It is trained on the same simulation time as our predictors, and its prediction for a start is the simulated time-averaged observable from that start's codeword. Beating it means higher explained variance, especially at later $t_w$, where a static landscape should fail if aging is real.

Third come scalar geometric descriptors fed to the same model: local contact counts, per-residue fluctuation, buried surface, and a density-based caging score analogous to the one in the fragility study. Fourth, the directional ablation removes the displacement direction from our model while keeping the distance, mirroring the ablation that hurt EIGNN on glasses. All learned models share the same hyperparameter search budget and data split.

On GlassBench-3D, the file reports no numerical correlation for EIGNN, only that it beat GlassMLP, CAGE and BOTAN at $T=0.44$ with about a fifth of BOTAN's parameters. We therefore report our reproduced numbers rather than claim to beat a published figure. For protein trapping prediction, no reported result was found on Lacuna.

Why not just…?

Why not just use the starting conformational state? This is the strongest threat, and the proposal is designed around it. If the basin label explains nearly all of $\phi$, the honest conclusion is this: protein non-ergodicity is coarse kinetic partitioning between a few long-lived states, and within-basin structure does not matter. That would still be a useful, quantitative result, because it would tell modelers that a few-state description suffices. The residual $\phi$ after conditioning on basin is the decisive number.

Why not just run a Markov state model or PESLA? A stationary Markov model with slow states can itself produce large $\phi$, so structural memory alone does not prove aging. The discriminator is waiting-time dependence. Under a stationary Markov model, the mapping from start state to outcome does not depend on $t_w$. Under aging, it drifts: $\phi$ and the per-state outcomes change with $t_w$ even after conditioning on the state. PESLA explicitly assumes a static, conservative landscape, so a $t_w$-dependent gap between it and the structure-based predictor is exactly what would favor aging.

Why not just simulate longer until things converge? That is the critique raised against protein aging, answered in Reply to: Insufficient evidence for ageing in protein dynamics. Our design does not rely on non-convergence. $\phi$ is defined at a fixed window $T$ chosen to match functional times, and the ssDNA control, run with the same protocol, sets the value expected when time averages converge. If SHP2's $\phi$ collapses toward the ssDNA value as $T$ grows within our budget, the skeptics gain ground, and we would report that.

Why not just train a generative trajectory model? Models such as ConfRover can emulate dynamics cheaply. ConfRover reports a 0.75 Pearson correlation on motion magnitude in ATLAS, against 0.56 for MDGen (Simultaneous Modeling of Protein Conformation and Dynamics via Autoregression, Generative Modeling of Molecular Dynamics Trajectories). These models were trained on 100-ns windows and do not separate structure from noise. We would include ConfRover rollouts from each start as an extra predictor. Whether it reproduces the within-start variance is itself an informative test of such emulators.

Decisive Checks

The first check is whether SHP2's within-basin $\phi$ at 100 ns is clearly above the ssDNA value, with bootstrap intervals that do not overlap. The second is whether $\phi$ changes monotonically with $t_w$ while the static-landscape baseline's accuracy degrades with $t_w$. The third is whether removing displacement direction lowers explained variance in proteins, as it does in glasses. The fourth is whether the receptive-field sweep saturates locally or only across domains, compared between K-Ras and PGK.

A null result on the first check falsifies the core claim. It would be a clean negative, showing that protein trapping in these simulations is not encoded in structure beyond the basin, and it would feed directly back into continuous-time random walk modeling.